Archive
A selection of postmortems published between 2025 and 2026 by teams running AI systems in production reveals repeated patterns: guardrail failures, silent model drift, hidden vendor dependency, and a collection of near-misses worth distilling.
Two years in, AI helps product discovery in one place above all: synthesizing interview transcripts. Generating hypotheses without real data has failed repeatedly, and simulated users produce systematic false positives about adoption. The practices that stick keep a human doing the critical analysis, because AI amplifies a good process and speeds a bad one toward failure.
The AI startup correction is already measurable: down rounds turned from anecdote into a visible statistical pattern from Q4 2025, and selective layoffs cluster in sales, research and operations at companies that over-hired. Survivors share a concrete problem, a concrete segment, and AI costs the business model can absorb. Thin wrappers over commercial models suffer most.
Claude Haiku 4.5, released by Anthropic on October 15, 2025, performs close to Sonnet 4 on structured tasks for roughly a third of the price. It carries a 200K-token context window and tool use on par with Sonnet 4.5. Pairing it as a filter ahead of Sonnet cuts total cost several times over.
Fully connecting the plant to global clouds now collides with European regulators and CFOs who no longer tolerate dependence on foreign providers; 2026 is the year we rethink where industrial data actually lives.
Carbon aware scheduling delivers savings in proportion to how much of your workload can move in time or geography. If under 20 percent of it is elastic, cluster-wide gains stay modest. Deferrable jobs do best, cutting carbon intensity 15 to 30 percent: nightly batch, model training, CI builds and tests, report generation, video rendering.
After the 2023-2024 hype cycle led by Apple Vision Pro, the 2025 valley of disillusionment, and the quiet but real consolidation of Meta Quest 3S and the WebXR stack, it is time to assess honestly where extended reality stands. What works, what has died, what is still alive.
WASI preview 3 was stabilized in late 2025 and adds what preview 2 lacked: native async, bidirectional streams, and inter-component concurrency with specified semantics. Portability and async are no longer a trade-off, since one async component runs unchanged on wasmtime, WasmEdge, JCO and the main commercial runtimes. API gateways and SaaS extension plugins are the strongest production cases.
Among the AI features in SRE dashboards, alert correlation is the one with demonstrated value: it groups dependent alerts from a single incident, cutting time to acknowledge and fatigue during a crisis. Automatic incident summaries help too. The real problem was never a shortage of information, it was separating signal from noise and correlating scattered symptoms.
